1. Bibliographic Information

1.1. Title

Medical image recognition and segmentation of pathological slices of gastric cancer based on Deeplab v3 + neural network

1.2. Authors

Jing Wang and Xiuping Liu. Both authors are affiliated with the Department of General Surgery, Shengjing Hospital of China Medical University, Liaoning, China. This affiliation suggests a clinical background, focusing on the practical application of AI in medicine.

1.3. Journal/Conference

The paper states it was published by "Elsevier B.V.". Elsevier is a major academic publishing company that hosts a wide range of reputable journals and conferences. However, the specific journal name is not mentioned in the provided text, which is unusual for a final published article. This might indicate it's from a conference proceeding published by Elsevier or a less-formally documented source.

1.4. Publication Year

2021. The paper was received on March 16, 2021, and accepted on May 24, 2021.

1.5. Abstract

The paper proposes an automatic segmentation model for cancerous regions in gastric cancer pathological slices using a Deeplab v3+ neural network. The objective is to enhance the efficiency and accuracy of diagnosis. The methodology involves using a dataset of 1240 pathological images to train a Deeplab v3+ model with a multi-scale input strategy. This model is then compared against SegNet and ICNet using metrics like sensitivity, specificity, accuracy, and the Dice coefficient. The results show that the proposed model achieves superior performance, with a Dice coefficient of 91.66%, which is over 12% higher than the baseline models. Additionally, the model is shown to be more computationally efficient. The conclusion is that the Deeplab v3+ based model is a promising tool for medical image analysis in gastric cancer diagnosis.

1.6. Original Source Link

/files/papers/691ab690110b75dcc59ae3f0/paper.pdf. This link points to a local PDF file. The paper includes a "© 2021 Elsevier B.V. All rights reserved." notice, indicating it has been officially published.

2. Executive Summary

2.1. Background & Motivation

The core problem this paper addresses is the significant challenge in diagnosing gastric cancer from histopathological slides. Gastric cancer is a major global health issue with high mortality rates, often due to late detection. Pathological examination of tissue slices is the "gold standard" for diagnosis, but this process is highly demanding. It requires expert pathologists to meticulously analyze vast amounts of visual data, leading to several challenges:

• High Workload & Fatigue: The sheer volume of data can cause pathologist fatigue, potentially compromising diagnostic accuracy.

• Shortage of Experts: There is a scarcity of trained pathologists, creating a bottleneck in the healthcare system.

• Subjectivity: Diagnosis can vary between different experts, affecting reliability.

  To address these gaps, the paper is motivated by the potential of deep learning, specifically Convolutional Neural Networks (CNNs), to automate the process of identifying and segmenting cancerous regions. The goal is to create an auxiliary diagnostic tool that can reduce the workload of pathologists, improve diagnostic efficiency, and increase the reliability of gastric cancer detection.

2.2. Main Contributions / Findings

The paper's primary contributions are:

1. A Specialized Segmentation Model: It proposes and validates a specific deep learning model, Deeplab v3+, enhanced with a multi-scale input strategy, for the task of segmenting cancerous regions in gastric cancer pathological images.
2. Comprehensive Performance Evaluation: The proposed model is rigorously compared against two other well-known segmentation architectures, SegNet and ICNet. The evaluation is based on multiple standard metrics (sensitivity, specificity, accuracy, Dice coefficient) as well as computational performance (memory usage, training time).
3. Demonstrated Superiority: The key finding is that the proposed Deeplab v3+ model significantly outperforms the baselines. It achieves a Dice coefficient of 91.66%, which is more than 12% higher than SegNet. Furthermore, it is shown to be more computationally efficient, requiring less video memory and shorter training time. This combination of high accuracy and efficiency makes it a practical solution for clinical settings.

3. Prerequisite Knowledge & Related Work

3.1. Foundational Concepts

• Histopathology: This is the microscopic examination of biological tissues to observe the appearance of diseased cells and tissues. In cancer diagnosis, a biopsy (a small tissue sample) is taken from the patient, thinly sliced, mounted on a glass slide, stained with special dyes to highlight different cellular components, and then examined under a microscope. The resulting image is a histopathological or pathological slice.
• Image Segmentation: This is a computer vision task that involves partitioning a digital image into multiple segments or sets of pixels. The goal is to simplify or change the representation of an image into something that is more meaningful and easier to analyze. In this paper, the task is semantic segmentation, where each pixel in the image is classified as belonging to a particular class (e.g., "cancerous tissue" or "healthy tissue"). This is more granular than image classification (assigning a single label to the whole image) or object detection (drawing a bounding box around objects).
• Convolutional Neural Networks (CNNs): A class of deep neural networks, most commonly applied to analyzing visual imagery. CNNs use a hierarchical pattern of layers to learn features from data. Key layers include:
  • Convolutional Layer: Applies filters (kernels) to an input image to create feature maps, detecting patterns like edges, textures, and shapes.
  • Pooling Layer: Reduces the spatial dimensions (width and height) of the feature maps, making the network more computationally efficient and robust to variations in the position of features.
  • Fully Connected Layer: A traditional neural network layer where every neuron is connected to every neuron in the previous layer, typically used at the end of a network for classification.
• Encoder-Decoder Architecture: A common architecture for semantic segmentation.
  • Encoder: This part of the network is typically a CNN that progressively downsamples the input image, extracting increasingly abstract and high-level features. As the spatial resolution decreases, the semantic information increases.
  • Decoder: This part takes the low-resolution, high-level feature maps from the encoder and progressively upsamples them to restore the original image resolution. The goal is to use the learned features to generate a pixel-wise segmentation map.

3.2. Previous Works

The paper builds upon several key deep learning models for semantic segmentation:

• Deeplab v3+: This model, proposed by Chen et al., is a state-of-the-art architecture for semantic segmentation. Its key innovations are:

  • Atrous (Dilated) Convolution: This is a type of convolution that allows the network to increase its receptive field (the area of the input image that influences a particular feature) without increasing the number of parameters or computational cost. It introduces gaps between the kernel weights, effectively "inflating" the kernel. This helps capture multi-scale context while maintaining spatial resolution.
  • Atrous Spatial Pyramid Pooling (ASPP): This module applies several parallel atrous convolutions with different "dilation rates" to the same feature map. By doing so, it can probe an incoming feature layer with filters at multiple scales, thus robustly segmenting objects of various sizes. The outputs are then concatenated to form a rich, multi-scale feature representation.
  • Encoder-Decoder Structure: Deeplab v3+ enhances the ASPP module by incorporating it into an encoder-decoder framework. This allows it to recover fine-grained object boundaries that might be lost during the downsampling in the encoder stage.

• SegNet: Developed by Badrinarayanan et al., SegNet is a classic encoder-decoder network. Its distinctive feature is the way it performs upsampling in the decoder. During the max-pooling operation in the encoder, the indices (locations) of the maximum values are stored. The decoder then uses these stored indices to place the values during its upsampling (unpooling) operation. This helps preserve high-frequency details and object boundaries more effectively than other upsampling methods like transposed convolution.

• ICNet (Image Cascade Network): Proposed by Zhao et al., ICNet is designed for real-time semantic segmentation. Its core idea is to use a multi-resolution input pipeline. It processes a low-resolution version of the image through the full semantic network to get a coarse prediction, then progressively refines this prediction using features from medium and high-resolution versions of the image. This cascade approach allows it to achieve a good balance between speed and accuracy.

3.3. Technological Evolution

The field of image segmentation has evolved significantly with the rise of deep learning. Initially, traditional computer vision techniques like thresholding and clustering were used. The advent of CNNs led to fully convolutional networks (FCNs) that could perform end-to-end pixel-level classification. Following FCNs, encoder-decoder architectures like SegNet and U-Net became popular for their ability to generate detailed segmentation maps. More recent advancements, as seen in the Deeplab series, have focused on capturing multi-scale context more effectively using techniques like atrous convolution and pyramid pooling. This paper positions itself at this modern stage by adopting Deeplab v3+, a highly advanced model, for a specific medical application.

3.4. Differentiation Analysis

The core innovation of this paper is not the invention of a new model but the application and empirical validation of an enhanced state-of-the-art model in a specific, high-impact domain.

• Compared to general Deeplab v3+: The authors propose using a multi-scale input strategy. While the paper does not detail this method, it generally involves feeding the network with versions of the input image at different scales or resolutions. This is a common technique to improve robustness to object size variations, and the paper's contribution is demonstrating its effectiveness for gastric cancer pathology.
• Compared to SegNet and ICNet: The paper differentiates its approach by showing that Deeplab v3+ is not only more accurate for this specific task but also more computationally efficient. This is a crucial finding for practical clinical deployment, where both performance and resource constraints are important. While ICNet is designed for speed, the results here suggest that Deeplab v3+ offers a better trade-off for this particular type of high-resolution medical image analysis.

4. Methodology

4.1. Principles

The core principle of the proposed method is to leverage the powerful feature extraction and multi-scale context aggregation capabilities of the Deeplab v3+ architecture for the semantic segmentation of cancerous regions in gastric cancer histopathology images. The model is trained in a supervised manner on a dataset of images where cancerous areas have been manually annotated. By learning the complex visual patterns (e.g., cell morphology, tissue structure, staining characteristics) that distinguish cancerous tissue from healthy tissue, the trained model can automatically generate a pixel-wise segmentation map for new, unseen images. The methodology is further enhanced by a multi-scale input strategy to improve the model's ability to handle variations in the size and appearance of cancerous lesions.

4.2. Core Methodology In-depth (Layer by Layer)

The overall workflow of the system is depicted in Figure 1, involving data input, enhancement, and segmentation using the trained model.

该图像是示意图，展示了胃癌病理切片图像的处理流程。从左侧的输入切片图像开始，经过数据增强后，进行胃切片图像分类、肿瘤目标检测和目标分割，最终输出分割结果。

The central components of the methodology are the data preprocessing and the deep learning model itself.

4.2.1. Data Preprocessing and Enhancement

Before feeding the images into the neural network, two main steps are performed:

1. Data Enhancement: To increase the diversity of the training data and prevent the model from overfitting, the authors apply several augmentation techniques. These include mirroring, flipping up and down, scaling, and rotation. This helps the model become more robust to variations in orientation and size of the cancerous regions.
2. Data Normalization: The pixel values of the input images are normalized to ensure the model converges faster and more stably during training. The paper uses a z-score normalization method. For each pixel, the new value $z$ is calculated from the original pixel value $x$. This is presented in Equation (1): $$z = { \frac { x - \omega } { s } }$$
   • Symbol Explanation:
     • $z$: The normalized pixel value, which will be the input to the network.
     • $x$: The original pixel value.
     • $\omega$: The mean value of all pixels across the entire training dataset.
     • $s$: The variance of all pixels across the entire training dataset.

4.2.2. The Deeplab v3+ Model

The paper uses Deeplab v3+ as its core segmentation model. The architecture, shown in Figure 2, consists of an encoder and a decoder.

该图像是Deeplab v3 +模型的网络结构示意图。左侧为输入的胃癌病理切片图像，右侧为分割结果，中央展示了编码器和解码器的详细结构，包括多个卷积层和上采样过程。

• Encoder: The encoder's role is to extract rich semantic information. It uses a backbone network (the paper mentions using an improved Xception module) to generate feature maps. At the end of the encoder, the Atrous Spatial Pyramid Pooling (ASPP) module is applied. ASPP uses multiple parallel atrous convolutions with different dilation rates to capture contextual information at various scales.
• Decoder: The decoder's goal is to recover fine-grained spatial information and object boundaries. It takes the feature maps from the ASPP module, upsamples them (by a factor of 4), and then concatenates them with low-level features from an earlier stage in the backbone network. These low-level features contain more detailed boundary information. Finally, a few more convolution layers and another upsampling step are applied to produce the final segmentation map, which has the same resolution as the input image.

4.2.3. Multi-Scale Input Strategy

The paper states that it proposes a "multi-scale input Deeplab v3+ network" and Table 3 shows that adding "multi-scale" improves the performance of all tested models. However, the paper does not provide any specific details on how this multi-scale input is implemented. It could involve creating an image pyramid and feeding each scale to the network, or other similar techniques, but this is a significant omission in the methodological description.

5. Experimental Setup

5.1. Datasets

The dataset used for this study consists of pathological slice images of gastric cancer. There is a discrepancy in the paper regarding the total number of images. The text mentions collecting 1340 images, but the sum of the training and test sets in Table 1 is 1240. The paper also states a 70% test set and 30% validation set split, which does not align perfectly with the numbers in the table (844 is ~68% of 1240, and 396 is ~32%).

The following are the results from Table 1 of the original paper:

The dataset was collected with the help of general surgeons at Shengjing Hospital, implying it consists of real patient data.

5.2. Evaluation Metrics

The performance of the segmentation models is evaluated using four metrics derived from a confusion matrix. A confusion matrix (Table 2) tabulates the performance of a classification model by comparing predicted labels to actual labels.

The four fundamental values are:

• True Positive (TP): Cancerous pixels correctly identified as cancerous.

• True Negative (TN): Non-cancerous pixels correctly identified as non-cancerous.

• False Positive (FP): Non-cancerous pixels incorrectly identified as cancerous.

• False Negative (FN): Cancerous pixels incorrectly identified as non-cancerous.

  Based on these values, the following metrics are used:

5.2.1. Sensitivity (Sen)

• Conceptual Definition: Also known as Recall or True Positive Rate, sensitivity measures the proportion of actual positive cases that were correctly identified. In this context, it answers the question: "Of all the actual cancerous pixels, what fraction did the model correctly identify?" High sensitivity is crucial in medical diagnosis to avoid missing diseases.
• Mathematical Formula: The formula is given by Equation (2): $$S e n = \frac { T P } { T P + F N }$$
• Symbol Explanation:
  • TP: True Positives
  • FN: False Negatives

5.2.2. Specificity (Spe)

• Conceptual Definition: Also known as the True Negative Rate, specificity measures the proportion of actual negative cases that were correctly identified. It answers: "Of all the actual non-cancerous pixels, what fraction did the model correctly identify?" High specificity is important to avoid false alarms or unnecessary treatments.
• Mathematical Formula: The paper provides Equation (3) for specificity: $$S p e = \frac { F P } { T P + T N }$$
• Symbol Explanation:
  • FP: False Positives
  • TP: True Positives
  • TN: True Negatives
• Critical Note: The formula provided in the paper for Specificity is non-standard and appears to be incorrect. The standard, universally accepted formula for Specificity is $Spe = \frac{TN}{TN + FP}$. The paper's formula does not correspond to any common evaluation metric. This is a significant error in the paper's methodology section.

5.2.3. Accuracy (Acc)

• Conceptual Definition: Accuracy measures the overall proportion of correct predictions (both positive and negative) among the total number of cases. It answers: "Overall, what fraction of pixels did the model classify correctly?"
• Mathematical Formula: The formula is given by Equation (4): $$A c c = { \frac { T P + T N } { T P + T N + F P + F N } }$$
• Symbol Explanation:
  • TP, TN, FP, FN: True Positives, True Negatives, False Positives, False Negatives.

5.2.4. Dice Coefficient (Dice)

• Conceptual Definition: The Dice coefficient is a statistic used to gauge the similarity of two samples. In image segmentation, it measures the overlap between the predicted segmentation mask and the ground truth (manually annotated) mask. A value of 1 indicates perfect overlap, while 0 indicates no overlap. It is particularly useful for evaluating segmentation tasks with imbalanced classes (e.g., small cancerous regions in a large image).
• Mathematical Formula: The paper provides Equation (5): $$D i c e = \frac { \sum _ { i = 1 } ^ { n } \sum _ { j = 1 } ^ { n } 2 t _ { i j } p _ { j i } } { \left( \sum _ { i = 1 } ^ { n } \sum _ { j = 1 } ^ { n } \left| t _ { i j } \right| ^ { 2 } \right) + \left( \sum _ { i = 1 } ^ { n } \sum _ { j = 1 } ^ { n } \left| p _ { i j } \right| ^ { 2 } \right) }$$
• Symbol Explanation:
  • $t$: The real label matrix (ground truth).
  • $p$: The predicted label matrix.
  • $t_{ij}$ and $p_{ji}$: Pixel values at specific coordinates.
• Critical Note: The formula provided for the Dice coefficient is highly unusual and likely contains typos. The use of transposed indices ($p_{ji}$) in the numerator is not standard. The standard formula for the Dice coefficient in the context of binary segmentation is: $ Dice = \frac{2 \times |A \cap B|}{|A| + |B|} $ where A is the predicted set of pixels and B is the ground truth set. For pixel-wise matrices, this is commonly written as: $ Dice = \frac{2 \sum_{i} t_i p_i}{\sum_{i} t_i^2 + \sum_{i} p_i^2} $ The paper's conceptual explanation matches the standard definition (twice the intersection over the sum of areas), but the mathematical formula presented is non-standard.

5.3. Baselines

The proposed Deeplab v3+ model is compared against two other deep learning-based segmentation models:

• SegNet: A representative encoder-decoder architecture known for its mechanism of using pooling indices to preserve boundary details.

• ICNet: A model designed for real-time segmentation that uses a multi-resolution image cascade to balance speed and accuracy.

  These baselines are appropriate as they represent different design philosophies in semantic segmentation, providing a good context for evaluating the performance of Deeplab v3+.

6. Results & Analysis

6.1. Core Results Analysis

The experimental results consistently demonstrate the superiority of the Deeplab v3+ model, especially when enhanced with the multi-scale input strategy.

Qualitative Results: Figure 5 provides a visual comparison of the segmentation results. The masks generated by Deeplab v3+ appear to align more closely with the cancerous regions in the original images compared to SegNet and ICNet, which seem to miss parts of the cancerous areas or produce less precise boundaries.

该图像是图表，展示了三个病例的胃癌病理切片在不同模型（Deeplab v3+、SegNet 和 ICNet）下的分割结果。左侧显示原始图像，右侧则分别是三种模型的分割效果，表明了不同算法在病变区域识别上的差异。

Quantitative Results: The primary quantitative results are summarized in Figure 6 and Table 3. The chart in Figure 6 shows a clear performance gap between Deeplab v3+ and the other two models across all four metrics.

该图像是图表，展示了不同模型（Deeplab v3+、SegNet、ICNet）在敏感性（Sen）、特异性（Spe）、准确率（Acc）及Dice系数（Dice）上的性能比较。从图中可以看出，Deeplab v3+在各项指标上均优于其他模型。

Table 3 provides a more detailed breakdown, including the impact of the multi-scale input strategy. From this table, two key conclusions can be drawn:

1. The multi-scale input strategy is highly effective. For all three models (Deeplab v3+, SegNet, and ICNet), adding multi-scale input improves performance across all metrics. For instance, the Dice score for Deeplab v3+ improves from 89.99% to 91.66%.

2. Deeplab v3+ is the superior architecture. Even without multi-scale input, Deeplab v3+ (Dice: 89.99%) outperforms the multi-scale versions of both SegNet (Dice: 82.01%) and ICNet (Dice: 80.33%). The final proposed model (Deeplab v3+ + multi-scale) achieves the highest scores, with a sensitivity of 91.45%, specificity of 92.31%, accuracy of 95.76%, and a Dice coefficient of 91.66%.

   Computational Efficiency: Table 4 analyzes the computational resource usage. The Deeplab v3+ model is shown to be the most efficient. It requires the least video memory (2.42 GB vs. >4 GB for others), has the highest GPU utilization (86.44%), and takes the shortest time to train (12.42 hours vs. >17 hours). This is a significant practical advantage for deployment in real-world clinical environments with limited hardware resources.

6.2. Data Presentation (Tables)

The following are the results from Table 3 of the original paper:

The following are the results from Table 4 of the original paper:

6.3. Ablation Studies / Parameter Analysis

The comparison between the base models and their +multi-scale versions in Table 3 serves as an ablation study for the multi-scale input component. The results clearly validate the effectiveness of this strategy, as its addition consistently boosts the performance of every model. For the best-performing model, Deeplab v3+, the multi-scale input provides an absolute improvement of 2.03% in sensitivity, 2.14% in specificity, 4.52% in accuracy, and 1.67% in the Dice coefficient. This confirms that the multi-scale strategy is a crucial part of the paper's proposed solution.

7. Conclusion & Reflections

7.1. Conclusion Summary

The paper successfully demonstrates that a deep learning model based on the Deeplab v3+ architecture, when enhanced with a multi-scale input strategy, can effectively and efficiently perform automatic segmentation of cancerous regions in gastric cancer pathological slices. The proposed model achieved a high Dice coefficient of 91.66%, significantly outperforming SegNet and ICNet by over 12% in key metrics. Furthermore, it proved to be more resource-efficient, requiring less memory and training time. The authors conclude that Deeplab v3+ is a valuable tool for assisting in the diagnosis of gastric cancer and warrants further promotion in medical image analysis.

7.2. Limitations & Future Work

The authors acknowledge several areas for future improvement:

• Handling Unlabeled Regions: The model currently treats some unlabeled positive regions (cancerous areas not marked in the ground truth) as negative samples during training. While a "directed distance field algorithm" (not explained in the paper) was used to mitigate this, the authors suggest that iterative methods to generate valid positive regions could further improve model performance.
• Incorporating Additional Information: To help the model learn more about cellular arrangements, the authors propose adding cell nucleus mask images as an additional input channel to the network in future experiments. This could provide the model with more granular information to improve segmentation accuracy.

7.3. Personal Insights & Critique

This paper presents a solid application of a state-of-the-art deep learning model to a critical medical problem, achieving impressive results. Its strengths lie in its clear problem definition, strong empirical evidence, and practical consideration of computational efficiency. However, the paper suffers from several notable weaknesses that affect its scientific rigor and reproducibility.

Strengths:

• Clinically Relevant: The research addresses a real-world bottleneck in cancer diagnosis, with the potential to have a significant positive impact on clinical workflows.
• Strong Empirical Results: The demonstrated superiority of the Deeplab v3+ model in both accuracy and efficiency is convincing and well-supported by the data presented.
• Good Baseline Comparison: The choice of SegNet and ICNet as baselines provides a solid context for evaluating the proposed model's performance.

Weaknesses and Areas for Improvement:

• Lack of Methodological Detail: The paper's most significant flaw is the complete absence of a description for its key innovation—the multi-scale input strategy. Without this detail, the work is not reproducible.

• Incorrect Formulas: The provided mathematical formulas for Specificity and the Dice Coefficient are non-standard and incorrect. This is a serious error that undermines the credibility of the paper's technical description, even if the actual calculations were likely performed correctly using standard libraries.

• Data Inconsistencies: The conflicting numbers regarding the dataset size (1340 vs. 1240) and the training/test split are signs of carelessness and poor proofreading, which reduces confidence in the overall quality of the research.

• Confusing Baseline Mention: The abstract mentions a comparison to "Faster-RCNN", which is an object detection model, not a segmentation model. It is not mentioned again in the experiments, suggesting this is another error.

  In conclusion, while the paper's findings are promising and highlight a valuable direction for AI in pathology, its technical execution and documentation are lacking in rigor. It serves as a good proof-of-concept, but for it to be a truly impactful scientific contribution, it would require a much more detailed and accurate methodological description.